A Refined Analysis of the Poisson Channel in the HighPhoton-Efficiency Regime
The MIT Faculty has made this article openly available. Please share
how this access benefits you. Your story matters.
Citation
Wang, Ligong, and Gregory W. Wornell. “A Refined Analysis of
the Poisson Channel in the High-Photon-Efficiency Regime.”
IEEE Trans. Inform. Theory 60, no. 7 (July 2014): 4299–4311.
As Published
http://dx.doi.org/10.1109/TIT.2014.2320718
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Version
Author's final manuscript
Accessed
Thu May 26 00:31:42 EDT 2016
Citable Link
http://hdl.handle.net/1721.1/91054
Terms of Use
Creative Commons Attribution-Noncommercial-Share Alike
Detailed Terms
http://creativecommons.org/licenses/by-nc-sa/4.0/
1
A refined analysis of the Poisson channel in the
high-photon-efficiency regime
arXiv:1401.5767v2 [cs.IT] 23 Apr 2014
Ligong Wang, Member, IEEE and Gregory W. Wornell, Fellow, IEEE
Abstract—We study the discrete-time Poisson channel under
the constraint that its average input power (in photons per
channel use) must not exceed some constant E . We consider the
wideband, high-photon-efficiency extreme where E approaches
zero, and where the channel’s “dark current” approaches zero
proportionally with E . Improving over a previously obtained firstorder capacity approximation, we derive a refined approximation,
which includes the exact characterization of the second-order
term, as well as an asymptotic characterization of the thirdorder term with respect to the dark current. We also show that
pulse-position modulation is nearly optimal in this regime.
Index Terms—Optical communication, Poisson channel, channel capacity, wideband, low SNR.
I. I NTRODUCTION
W
E consider the discrete-time memoryless Poisson channel whose input x is in the set R+
0 of nonnegative reals
and whose output y is in the set Z+
of
nonnegative integers.
0
Conditional on the input X = x, the output Y has a Poisson
distribution of mean (λ + x), where λ ≥ 0 is called the “dark
current” and is a constant that does not depend on the input X.
We denote the Poisson distribution of mean ξ by Poisξ (·) so
Poisξ (y) = e−ξ
ξy
,
y!
y ∈ Z+
0.
(1)
output y is the actual number of photons that are detected in
the pulse duration. The dark current λ is the average number
of extraneous counts that appear in y. We note that, although
the name “dark current” is traditionally used in the Poissonchannel literature, these extraneous counts usually include both
the detector’s “dark clicks” (which are where the name “dark
current” comes from) and photons in background radiation.
We impose an average-power constraint2 on the input
E[X] ≤ E
for some E > 0.
In applications like free-space optical communications, the
cost of producing and successfully transmitting photons is
high, hence high photon-information efficiency—amount of
information transmitted per photon, which we henceforth call
simply “photon efficiency”—is desirable. As we later demonstrate, this can be achieved in the wideband regime, where
the pulse duration of the input approaches zero and, assuming
that the continuous-time average input power is fixed, where E
approaches zero proportionally with the pulse duration. Note
that in this regime the average number of detected background
photons or dark clicks also tends to zero proportionally with
the pulse duration. Hence we have the linear relation
With this notation the channel law W (·|·) is
W (y|x) = Poisλ+x (y),
x∈
R+
0 ,y
λ = cE,
∈
Z+
0.
(2)
This channel models pulse-amplitude modulated optical
communication where the transmitter sends light signals in
coherent states (which are usually produced using laser devices), and where the receiver employs direct detection (i.e.,
photon counting) [1]. The channel input x describes the
expected number of signal photons (i.e., photons that come
from the input light signal rather than noise) to be detected in
the pulse duration, and is proportional to the light signal’s
intensity,1 the pulse duration, the channel’s transmissivity,
and the detector’s efficiency; see [1] for details. The channel
This work was supported in part by the DARPA InPho program under
Contract No. HR0011-10-C-0159, and by AFOSR under Grant No. FA955011-1-0183. This paper was presented in part at the 2012 IEEE Information
Theory Workshop (ITW) in Lausanne, Switzerland, and will be presented in
part at the 2014 IEEE International Symposium on Information Theory (ISIT)
in Honolulu, HI, USA.
L. Wang is with the Massachusetts Institute of Technology, Cambridge, MA
02139 USA (e-mail: wlg@mit.edu).
G.W. Wornell is with the Massachusetts Institute of Technology, Cambridge,
MA 02139 USA (e-mail: gww@mit.edu).
1 We assume that the pulses are square, which is usually the case in practice.
If they are not square, the light signal’s intensity should be averaged over the
pulse duration.
(3)
(4)
where c is some nonnegative constant that does not change
with E. In practice, asymptotic results in this regime are useful
in scenarios where E is small and where λ is comparable to
or much smaller than E. Scenarios where E is small but λ
is much larger is better captured by the model where λ stays
constant while E tends to zero; see [2, Proposition 2].
We denote the capacity (in nats per channel use) of the
channel (2) under power constraint (3) with dark current (4)
by C(E, c), then
C(E, c) = sup I(X; Y ),
(5)
E[X]≤E
where the mutual information is computed from the channel
law (2) and is maximized over input distributions satisfying (3), with dark current λ given by (4). As we shall
see, our results on the asymptotic behavior of C(E, c) hold
irrespectively of whether a peak-power constraint
X ≤A
with probability 1
(6)
2 Here “power” is in discrete time, means expected number of detected
signal photons per channel use, and is proportional to the continuous-time
physical power times the pulse duration.
2
is imposed or not, as long as A is positive and does not
approach zero together with E.
We now formally define the maximum achievable photon
efficiency CPE (E, c), where the subscript notation services
a reminder that this quantity is photon efficiency and not
capacity:
C(E, c)
CPE (E, c) ,
.
(7)
E
Various capacity results for the discrete-time Poisson channel have been obtained [2]–[6]. Among them, [2, Proposition
1] considers the same scenario as the present paper and asserts
that
C(E, c)
lim
= 1, c ∈ [0, ∞).
(8)
E↓0 E log 1
E
In other words, the maximum photon efficiency satisfies
1
1
, c ∈ [0, ∞).
(9)
CPE (E, c) = log + o log
E
E
Furthermore, the proof of [2, Proposition 1] shows that the
limit (8) is achievable using on-off keying. In the Appendix
we provide a proof for the converse part of (8) that is much
simpler than the original proof in [2].3
The approximation in (9) can be compared to the maximum
photon efficiency achievable on the pure-loss bosonic channel
[7]. This is a quantum optical communication channel that
attenuates the input optical signal, but that does not add any
noise to it. The transmitter and the receiver of this channel
may employ any structure permitted by quantum physics.
When the transmitter is restricted to sending coherent optical
states, and when the receiver is restricted to ideal direct
detection (with no dark counts), the pure-loss bosonic channel
becomes the Poisson channel (2) with λ = 0. We denote
by CPE-bosonic (E) the maximum photon efficiency of the pureloss bosonic channel under an average-power constraint that
is equivalent to (3). The value of CPE-bosonic (E) can be easily
computed using the explicit capacity formula [7, Eq. (4)],
which yields
1
+ 1 + o(1).
(10)
E
Comparing (9) and (10) shows the following:
• For the pure-loss bosonic channel in the wideband
regime, coherent-state inputs and direct detection are
optimal up to the first-order term in photon efficiency
(or, equivalently, in capacity). For example, they achieve
infinite capacity per unit cost [8], [9].
• The dark current does not affect this first-order term.
In the present paper, we refine the analysis in [2] in two
aspects. First, we provide a more accurate approximation for
CPE (E, c) that contains higher-order terms, and that reflects the
influence of the dark current, i.e., of the constant c. Second, we
identify near-optimal modulation schemes that facilitate code
design for this channel.
CPE-bosonic (E) = log
3 For the achievability part of (8), we note that its proof can be simplified
by letting the decoder ignore multiple photons, rather than considering all
possible values of Y as in [2]. This can be seen via the achievability proofs
in the current paper, which do ignore multiple photons at the receiver, and
which yield stronger results than those in [2].
Some progress has been made in improving the approximation (9). It is noticed in [10] that the photon efficiency
achievable on the Poisson channel (2) with λ = 0 may be of
the form
1
1
(11)
log − log log + O(1),
E
E
which means that restriction to coherent-state inputs and direct
detection may induce a loss in the second-order term in
the photon efficiency on the pure-loss bosonic channel. For
practical values of E, this second-order term can be significant.
For example, for E = 10−5 , the difference between the firstorder term in (9) and the first- and second-order terms in (11)
is about 20%.
The analysis in [10] (whose main focus is not on the
Poisson channel itself) is based on certain assumptions on the
input distribution.4 It is therefore unclear from [10] if (11)
is the maximum photon efficiency achievable on the Poisson
channel (2) with λ = 0 subject to constraint (3) alone, i.e.,
if (11) is the correct expression for CPE (E, 0). In the present
paper, we prove that this is indeed the case.
Recent works such as [10], [11] often ignore the dark current
in the Poisson channel. It has been unknown to us whether
the dark current, i.e., whether the constant c influences the
second term in (11) and, if not, whether it influences the next
term in photon efficiency. We show that the constant c does
not influence the second-order term in CPE (E, c), but does
influence the third-order, constant term.
It has long been known that that infinite photon efficiency
on the Poisson channel with zero dark current can be achieved
using pulse-position modulation (PPM) combined with an
outer code [3], [4]. Further, [11] observes that PPM can
achieve (11) on such a channel. PPM greatly simplifies the
coding task for this channel, since one can easily apply existing
codes, such as a Reed-Solomon code, to the PPM “super
symbols”; while the on-off keying scheme that achieves (11)
has a highly skewed input distribution and is hence difficult
to code. The question then arises: how useful is PPM when
there is a positive dark current? This question has two parts.
First, is PPM still near optimal in terms of capacity (or photon
efficiency) when there is a positive dark current? Second, does
PPM still simplify coding when there is a dark current? We
answer the first part of the question in the affirmative in our
main results. We cannot fully answer the second question in
this paper, but we shall discuss it in the concluding remarks
in Section VII.
The rest of this paper is arranged as follows. We introduce
some notation and formally define PPM in Section II. We state
and discuss our main results in Section III. We then prove the
achievability parts of these results in Sections IV and V, and
prove the converse parts in Section VI. We conclude the paper
with numerical comparison of the bounds and some remarks
in Section VII.
II. N OTATION AND PPM
We usually use a lower-case letter like x to denote a
constant, and an upper-case letter like X to denote a random
variable.
4 According
to conversation with the authors.
3
We use natural logarithms, and measure information in nats.
We use the usual o(·) and O(·) notation to describe behaviors of functions of E in the limit where E approaches zero
with other parameters, if any, held fixed. Specifically, given
a reference function f (·) (which might be the constant 1), a
function described as o(f (E)) satisfies
lim
E↓0
o(f (E))
= 0,
f (E)
and a function described as O(f (E)) satisfies
O(f (E)) < ∞.
lim E↓0
f (E) (12)
(13)
We emphasize that, in particular, we do not use o(·) and O(·)
to describe how functions behave with respect to c.
We adopt the convention
0 log 0 = 0.
(14)
We next formally describe what we mean by PPM. On the
transmitter side:
• The channel uses are divided into frames of equal lengths;
• In each frame, there is only one channel input that is
positive, which we call the “pulse”, while all the other
inputs are zeros;
• The pulses in all frames have the same amplitude.
On the receiver side, we distinguish between two cases, which
we call simple PPM and soft-decision PPM, respectively. In
simple PPM, the receiver records at most one pulse in each
frame; if more than one pulse is detected in a frame, then that
frame is recorded as an “erasure”, as if no pulse is detected
at all. In soft-decision PPM, the receiver records up to two
pulses in each frame; frames containing no or more than two
detected pulses are treated as erasures.
III. M AIN R ESULTS
AND
D ISCUSSIONS
The following theorem provides an approximation for
the photon efficiency CPE (E, c) in the high-photon-efficiency
regime.
Theorem 1. The maximum photon efficiency CPE (E, c) achievable on the Poisson channel (2) subject to constraint (3) and
with dark current (4), which we define in (7), satisfies
1
1
− log log + O(1), c ∈ [0, ∞). (15)
E
E
Furthermore, denote
1
1
(16)
K(E, c) , CPE (E, c) − log + log log ,
E
E
then
K(E, c)
K(E, c)
lim lim
= lim lim
= −1.
(17)
c→∞ E↓0 log c
c→∞ E↓0 log c
CPE (E, c) = log
Our second theorem shows that PPM is nearly optimal in
the regime of interest.
Theorem 2. The asymptotic expression on the right-hand side
of (15) is achievable with simple PPM. The limits in (17) are
achievable with soft-decision PPM.
The achievability part of Theorem 1 follows from Theorem 2. To prove the latter, we need to show two things: first,
that the largest photon efficiency that is achievable with simple
PPM, which we henceforth denote by CPE-PPM (E, c), satisfies
CPE-PPM (E, c) ≥ log
1
1
−log log +O(1),
E
E
c ∈ [0, ∞); (18)
and second, that the maximum photon efficiency that is achievable with soft-decision PPM, which we henceforth denote by
CPE-PPM(SD) (E, c), satisfies
CPE-PPM(SD) (E, c) − log
lim lim
log c
c→∞ E↓0
1
1
+ log log
E
E ≥ −1. (19)
To prove the converse part of Theorem 1, we note that the
capacity of the channel is monotonically decreasing in c [2],
and hence so is the photon efficiency. It thus suffices to show
two things: first, that, in the absence of dark current, the largest
photon efficiency achievable with any scheme satisfies
CPE (E, 0) ≤ log
1
1
− log log + O(1);
E
E
(20)
and second, that
1
1
+ log log
E
E ≤ −1.
log c
CPE (E, c) − log
lim lim
c→∞ E↓0
(21)
We prove (18) in Section IV, (19) in Section V, and (20)
and (21) in Section VI.
To better understand the capacity results in Theorem 1, we
make the following remarks.
• Choosing c = 0 in (15) confirms that (11) is the correct
asymptotic expression up to the second-order term for
CPE (E, 0). Compared to (10) this means that, on the
pure-loss bosonic channel, for small E, restricting the
receiver to using direct detection induces a loss in photon
efficiency of about log log E1 nats per photon. Note that
the capacity of the pure-loss bosonic channel can be
achieved using coherent input states only [7], so this loss
is indeed due to direct detection, and not due to coherent
input states.
Attempts to overcome this loss by employing other feasible detection techniques have so far been unsuccessful
[10], [12].
• The expression on the right-hand side of (15) does not
depend on c, so the value of c affects neither the firstorder term nor the second-order term in CPE (E, c). In
particular, these two terms do not depend on whether c
is zero or positive.
• The first term in CPE (E, c) that c does affect is the third,
constant term. Indeed, though we have not given an exact
expression for the constant term, the asymptotic property
(17) shows that c affects the constant term in such a
way that, for large c, the constant term is approximately
− log c.
• Theorem 1 suggests the approximation
CPE (E, c) ≈ log
1
1
− log log − log c.
E
E
(22)
4
The terms constituting the approximation error, roughly
speaking, are either vanishing for small E, or small
compared to log c for large c. Note that if we fix the dark
current, i.e., if we fix the product cE, then the first and
third terms on the right-hand side of (22) cancel. This is
intuitively in agreement with [2, Proposition 2], which
asserts that, for fixed dark current, photon efficiency
scales like some constant times log log(1/E) and hence
not like log(1/E). We note, however, that (17) cannot be
derived directly using (15) together with [2, Proposition
2], because we cannot change the order of the limits. In
(17) we do not let E tend to zero and c tend to infinity
simultaneously. Instead, we first let E tend to zero to close
down onto the constant term in CPE (E, c) with respect to
E, and then let c tend to infinity to study the asymptotic
behavior of this constant term with respect to c.
• The approximation (22) is good for large c, but diverges
as c tends to zero. We hence need a better approximation
for the constant term, which behaves like − log c for
large c, but which does not diverge for small c. As both
the nonasymptotic bounds and the numerical simulations
we show later will suggest, − log(1 + c) is a good
approximation:
1
1
(23)
CPE (E, c) ≈ log − log log − log(1 + c).
E
E
For modulation and coding considerations, we make some
remarks following Theorem 2.
• Theorem 2 shows that PPM is optimal up to the secondorder term in photon efficiency when c = 0. Hence,
compared to the restriction to the receiver using direct
detection, the further restriction to PPM induces only a
small additional loss in photon efficiency.
• Furthermore, for the third-order, constant term, (softdecision) PPM is also not far from optimal, in the sense
that it achieves the optimal asymptotic behavior of this
term for large c.
• In Section IV we show that
•
CPE-PPM (E, c)
1
3
1
≥ log − log log − c − log(1 + c) − + o(1). (24)
E
E
2
A careful analysis will confirm that the bound (24) is
tight in the regime of interest, in the sense that simple
PPM cannot achieve a constant term that is better than
linear in c (while being second-order optimal). This is
in contrast to soft-decision PPM, which can achieve a
constant term that is logarithmic in c. In particular, simple
PPM is clearly not third-order optimal.5
In the PPM schemes that achieve (18) and (19), which we
describe in Sections
IV and V, the pulse has amplitude
1/ log(1/E) , which depends on E, and which tends to
zero as E tends to zero. An on-off keying scheme having
the same amplitude for its “on” signals can also achieve
5 A scheme between simple and soft-decision PPM is the following. When
detecting multiple pulses in a frame, the receiver randomly selects and records
one of the positions (possibly together with a “quality” flag). This scheme
outperforms simple PPM in photon efficiency, but its third-order term is still
linear in c: it scales like −c/2 instead of −c.
•
these lower bounds, because the dependence between
the channel inputs introduced by PPM can only reduce
the total mutual information between channel inputs and
outputs. These (PPM and on-off keying schemes) are
different from the on-off keying scheme used in [2] where
the “on” signal has a fixed amplitude that does not depend
on E. The latter on-off keying scheme, as well as any
PPM scheme with a fixed pulse amplitude (with respect
to E), is not second-order optimal on the Poisson channel.
Because in the PPM schemes to achieve (18) and (19) the
pulse tends to zero as E tends to zero, we know that, as
claimed in Section I, both our theorems hold if a constant
(i.e., not approaching zero together with E) peak-power
constraint as in (6) is imposed on X in addition to (3).
We next proceed to prove our main results, and leave further
discussions to Section VII.
IV. P ROOF
OF THE
S IMPLE -PPM L OWER B OUND (18)
Assume that E is small enough so that
E log
1
<1
E
E < e−c .
(25a)
(25b)
Consider the following simple PPM scheme:
Scheme 1.
•
The channel uses are divided into frames of length b, so
each frame contains b input symbols x1 , . . . , xb and b
corresponding output symbols y1 , . . . , yb . We set
1
.
(26)
b=
E log E1
•
Within each length-b frame, there is always one input that
equals η, and all the other (b − 1) inputs are zeros. Each
frame is then fully specified by the position of its unique
nonzero symbol, i.e., its pulse position. We consider each
frame as a “super input symbol” x̃ that takes value in
{1, . . . , b}. Here x̃ = i, i ∈ {1, . . . , b}, means
xi = η
xj = 0,
(27a)
j ∈ {1, . . . , b}, j 6= i.
(27b)
To meet the average-power constraint (3) with equality,
we require
η = bE.
•
(28)
The b output symbols y1 , . . . , yb are mapped to one
“super output symbol” ỹ that takes value in {1, . . . , b, ?}
in the following way: ỹ = i, i ∈ {1, . . . , b}, if yi is the
unique nonzero term in {y1 , . . . , yb }; and ỹ = ? if there
is more than one or no nonzero term in {y1 , . . . , yb }.
We have the following lower bound on the photon efficiency
achieved by the above scheme.
5
Proposition 1. Under the conditions (25), Scheme 1 achieves
photon efficiency
CPE-PPM (E, c)
η
cE
≥ 1−
log b − cη log b − 1 +
log(1 + c)
2
η
η o
cE n
log(1 − cη) + log 1 −
+ 1+
η
2
cE
(29)
−1−
η
W̃ (?|i) = 1 − p0 − (b − 1)p1 ,
Denote the capacity of this super channel by C̃(E, c, b, η), then
C̃(E, c, b, η) = max I(X̃; Ỹ ).
and simplify (29) to obtain
CPE-PPM (E, c)
1
1
3
≥ log − log log − c − log(1 + c) − + o(1) (31)
E
E
2
1
1
(32)
= log − log log + O(1).
E
E
When proving Proposition 1, as well as Propositions 2
and 3, we frequently use the inequalities summarized in the
following lemma.
Lemma 1. For all a ≥ 0,
1−e
−a
(33)
≤ a,
(34)
1
1 − e−a ≥ a − a2 .
(35)
2
Proof of Proposition 1: We compute the transition matrix
of the PPM “super channel” as follows:
W̃ (i|i) = Pr[Ỹ = i|X̃ = i]
(36)
= Pr [ Yi ≥ 1| Xi = η]
b
Y
Pr [ Yk = 0 |Xk = 0] (37)
k=1
k6=i
b−1
= 1 − W (0|η) W (0|0)
(39)
=e
, p0 ,
−e
i ∈ {1, . . . , b};
(40)
(41)
W̃ (j|i) = Pr[Ỹ = j|X̃ = i]
= Pr [ Yi = 0| Xi = η] Pr [ Yj ≥ 1| Xj = 0]
b
Y
·
Pr [ Yk = 0| Xk = 0]
(42)
= (1 − e
−η−cE
)e
−(b−1)cE
−(b−1)cE
−η−bcE
(38)
(43)
=e
−η−cE
(1 − e
−cE
)e
−(b−2)cE
= e−η−(b−1)cE − e−η−bcE
, p1 ,
i, j ∈ {1, . . . , b}, i 6= j;
Note that the total input power (i.e., expected number of
detected signal photons) in each frame equals η. Therefore
we have the following lower bound on CPE-PPM (E, c):
C̃(E, c, b, η)
.
(51)
η
It can be easily verified that the optimal input distribution
for (50) is the uniform distribution
1
PX̃ (i) = , i ∈ {1, . . . , b},
(52)
b
which induces the following marginal distribution on Ỹ :
CPE-PPM (E, c) ≥
p0 + (b − 1)p1
, i ∈ {1, . . . , b},
b
PỸ (?) = 1 − p0 − (b − 1)p1 .
PỸ (i) =
(44)
(45)
(46)
(47)
(53a)
(53b)
We can now use the above joint distribution on (X̃, Ỹ ) to
lower-bound C̃(E, c, b, η) as follows:
C̃(E, c, b, η)
= I(X̃; Ỹ )
= H(Ỹ ) − H(Ỹ |X̃)
(54)
(55)
1
1 − p0 − (b − 1)p1
b
+ (p0 + (b − 1)p1 ) log
p0 + (b − 1)p1
1
− (1 − p0 − (b − 1)p1 ) log
1 − p0 − (b − 1)p1
1
1
− (b − 1)p1 log
(56)
− p0 log
p0
p1
bp0
bp1
= p0 log
+ (b − 1)p1 log
p0 + (b − 1)p1
p0 + (b − 1)p1
(57)
p0 + (b − 1)p1
= p0 log b − p0 log
p
|
{z 0 }
= (1 − p0 − (b − 1)p1 ) log
≤log 1+
bp1
p0
p0 + (b − 1)p1
− (b − 1)p1 log
bp
{z 1
}
| =log 1+
k=1
k∈{i,j}
/
b−2
= W (0|η) 1 − W (0|0) W (0|0)
(50)
PX̃
Proof of (18) using Proposition 1: We plug (26) and
(28) into (29), note that
1
1
=
+ O(1)
(30)
1
E log E
E log E1
(48)
We now have, irrespective of the distribution of X̃,
1
1
+ (b − 1)p1 log
H(Ỹ |X̃) = p0 log
p0
p1
1
.
+ 1 − p0 − (b − 1)p1 log
1 − p0 − (b − 1)p1
(49)
where b and η are given in (26) and (28), respectively.
log(1 + a) ≤ a,
i ∈ {1, . . . , b}.
p0 −p1
bp1
≥ p0 log b − p0 log 1 +
≥ p0 log b − p0 log 1 +
≤
(58)
p0 −p1
bp1
bp1
p0
−
bp1
p0
− p0 .
b−1
(p0 − p1 )
b } | {z }
| {z
≤1
(59)
≤p0
(60)
6
Next note that p0 can be upper-bounded as
p0 = e−(b−1)cE − e−η−bcE
=e
−(b−1)cE
1−e
−η−cE
≤ 1 − e−η−cE
≤ η + cE,
(61)
(62)
(63)
(64)
where the last step follows from (34). It can also be lowerbounded as
p0 = e−(b−1)cE − e−η−bcE
≥ e−bcE − e−η−bcE
(65)
(66)
− e−η )
= e|−bcE
{z } (1
| {z }
(67)
≥1−bcE
≥η− 12 η 2
1
≥ (1 − bcE) η − η 2 ,
2
(68)
where the last inequality follows from (34) and (35), and from
the fact that the first multiplicand on its right-hand side is (in
fact, both multiplicands are) nonnegative for E satisfying (25)
and other parameters chosen accordingly. Also note that p1
can be upper-bounded as
p1 = e−η−(b−1)cE − e−η−bcE
e−cE
= |e−η−(b−1)cE
{z
} |1 −{z
}
≤1
(69)
(70)
≤cE
≤ cE.
(71)
Using (64), (68) and (71) we can continue the chain of
inequalities (60) to further lower-bound C̃(E, c, b, η) as
C̃(E, c, b, η)
1
≥ (1 − bcE) η − η 2 log b
2
− (η + cE) log 1 +
− η − cE
bcE
(1 − bcE) η − 21 η 2
!
Here, (73) follows from the fact that, for all α ≥ 0 and β ≥ 1,
log(1 + αβ) ≤ log(β + αβ) = log β + log(1 + α),
(75)
with the choices
β=
1
,
(1 − bcE) 1 − 21 η
V. P ROOF
OF THE
S OFT-D ECISION -PPM L OWER
B OUND (19)
Consider the following soft-decision PPM scheme:
Scheme 2. The transmitter performs the same PPM as in
Scheme 1. The receiver maps the b output symbols
to a
“super symbol”
ŷ
that
takes
value
in
{1,
.
.
.
,
b,
?}
∪
{i, j} ⊂
{1, . . . , b} . The mapping rule is as follows. Take ŷ = i if yi
is the unique nonzero term in {y1 , . . . , yb }; take ŷ = {i, j} if
yi and yj are the only two nonzero terms in {y1 , . . . , yb }; if
there are more than two or no nonzero term in {y1 , . . . , yb },
take ŷ = ?.
Proposition 2. Under the conditions (25), Scheme 2 achieves
photon efficiency
CPE-PPM(SD) (E, c)
cE
η
log b − cη log b − 1 +
log(1 + c)
≥ 1−
2
η
η o
cE
cE n
log(1 − cη) + log 1 −
−1−
+ 1+
η
2
η
c2 E 2
b
η
(b − 1) cE −
(1 − cη) log
+ 1−
2
2
2
c2
− η − c(η + cE)
(78)
2
where b and η are given in (26) and (28), respectively.
(72)
bcE
1 2
≥ (1 − bcE) η − η log b − (η + cE) log 1 +
2
η
n
η o
+ (η + cE) log(1 − bcE) + log 1 −
2
− η − cE
(73)
1 2
bcE
≥ η − η log b − bcEη log b − (η + cE) log 1 +
2
η
n
η o
+ (η + cE) log(1 − bcE) + log 1 −
2
− η − cE.
(74)
bcE
,
α=
η
which indeed satisfy α ≥ 0 and β ≥ 1; and (74) follows by
dropping the term 21 bcEη 2 log b, which is positive.
Combining (28), (51), and (74) yields (29).
Proof of (19) using Proposition 2: Simplifying (78) we
obtain
1
1
lim CPE-PPM(SD) (E, c) − log + log log
E
E
E↓0
3
≥ − log(1 + c) − . (79)
2
This immediately yields (19).
Proof of Proposition 2: We compute the transition matrix
of the super channel that results from Scheme 2. We first note
Ŵ (i|i) = p0 ,
Ŵ (j|i) = p1 ,
where p0 and p1 are given in (41) and (47), respectively. For
the remaining elements of the transition matrix we have
Ŵ {i, j}i = Pr[Yi ≥ 1|Xi = η] Pr[Yj ≥ 1|Xj = 0]
·
(76)
(77)
i, j ∈ {1, . . . , b}, i 6= j,
(80)
(81)
b
Y
Pr[Yℓ = 0|Xℓ = 0]
(82)
ℓ=1
ℓ∈{i,j}
/
= (1 − e−η−cE )(1 − e−cE ) e−(b−2)cE
, p2 , {i, j} ⊂ {1, . . . , b};
(83)
(84)
7
Ŵ {j, k}i = Pr[Yi = 0|Xi = η] Pr[Yj ≥ 1|Xj = 0]
· Pr[Yk ≥ 1|Xk = 0]
b
Y
·
Pr[Yℓ = 0|Xℓ = 0]
(85)
= e−η−cE (1 − e−cE )2 e−(b−3)cE
, p3 , {i, j, k} ⊂ {1, . . . , b};
(86)
(87)
ℓ=1
ℓ∈{i,j,k}
/
Ŵ (?|i) = 1 − p0 − (b − 1)p1 − (b − 1)p2
(b − 1)(b − 2)
p3
−
2
, p4 , i ∈ {1, . . . , b}.
(88)
(89)
bp2
2p2 + (b − 2)p3
b
= (b − 1)p2 log − (b − 1)p2
2
(b − 1)p2 log
log
|
p2 + b−2
2 p3
p2
{z
}
(97)
(b−2)p
(b−2)p
=log 1+ 2p 3 ≤ 2p 3
b (b − 1)(b − 2)
p3
≥ (b − 1)p2 log −
2
2
2
b b
≥ (b − 1)p2 log − p3 .
2
2
2
2
(98)
(99)
We lower-bound the fourth term on the right-hand side of (96)
as
We now have, irrespective of the distribution of X̃,
1
1
+ (b − 1)p1 log
p0
p1
1
(b − 1)(b − 2)
1
+ (b − 1)p2 log
+
p3 log
p2
2
p3
1
+ p4 log .
(90)
p4
H(Ŷ |X̃) = p0 log
Choosing a uniform X̃ (which is optimal as in Section IV),
we have the following distribution on Ŷ
p0 + (b − 1)p1
, i ∈ {1, . . . , b};
(91)
b
2p2 + (b − 2)p3
{i, j} =
, {i, j} ⊂ {1, . . . , b}; (92)
b
PŶ (?) = p4 .
(93)
PŶ (i) =
PŶ
side of (96) as
Therefore
(b − 1)(b − 2)
bp3
p3 log
2
2p2 + (b − 2)p3
2p2 + (b − 2)p3
(b − 1)(b − 2)
p3
log
=−
2
bp
|
{z 3
}
(100)
2(p2 −p3 )
2p
2(p2 −p3 )
=log 1+
≤ bp2
≤
bp
bp
3
3
3
(b − 1)(b − 2) 2p2
≥−
·
2
b
≥ −bp2 .
(101)
(102)
Using (34) and (35) we have the upper bound on p2
p2 = (1 − e−η−cE ) (1 − e−cE ) e|−(b−2)cE
{z }
{z
} | {z }
|
≤η+cE
(103)
≤1
≤cE
≤ (η + cE)cE,
(104)
the lower bound on p2
H(Ŷ )
b
= p0 + (b − 1)p1 log
p0 + (b − 1)p1
b
(b − 1)(b − 2)
p3 log
+ (b − 1)p2 +
2
2p2 + (b − 2)p3
1
(94)
+ p4 log .
p4
p2 = (1 − e−η−cE) (1 − e−cE) |e−(b−2)cE
{z }
|
{z
} | {z }
≥1−e−η ≥η− η
2
≥cE− c
2 E2
2
≥1−bcE
2
η2
c2 E 2
≥ η−
cE −
(1 − bcE),
2
2
p3 = |e−η−cE
e−cE ) e|−(b−3)cE
{z } |(1 −{z
{z }
}
≤1
I(X̃; Ŷ ) = H(Ŷ ) − H(Ŷ |X̃)
bp0
= p0 log
p0 + (b − 1)p1
(95)
bp1
+ (b − 1)p1 log
p0 + (b − 1)p1
bp2
+ (b − 1)p2 log
2p2 + (b − 2)p3
(b − 1)(b − 2)
bp3
+
. (96)
p3 log
2
2p2 + (b − 2)p3
At this point, note that the first two summands on the righthand side of (96) constitute I(X̃; Ỹ ), which we analyzed in
Section IV. It remains to find lower bounds on the last two
summands. We lower-bound the third term on the right-hand
(106)
and the upper bound on p3
2
Using (90) and (94) we have
(105)
2 2
≤c E .
≤cE
(107)
≤1
(108)
From (96), (99), (102), (104), (106), and (108) we obtain
a lower bound on the additional mutual information that
is gained by considering output frames with two detection
positions:
I(X̃; Ŷ ) − I(X̃; Ỹ )
b b2
(109)
≥ (b − 1)p2 log − p3 − bp2
2 2
c2 E 2
b
η2
cE −
(1 − bcE) log
≥ (b − 1) η −
2
2
2
b2
− c2 E 2 − b(η + cE)cE.
(110)
2
8
Dividing the above by η and plugging in (28) we obtain a
lower bound on the additional photon efficiency:
I(X̃; Ŷ ) − I(X̃; Ỹ )
η
c2 E 2
b
η
(b − 1) cE −
(1 − cη) log
≥ 1−
2
2
2
c2
− η − c(η + cE).
(111)
2
Adding the right-hand side of (111) to the right-hand side of
(29) yields (78).
VI. P ROOF
OF THE
U PPER B OUNDS (20) AND (21)
Proposition 3. Assume that E is small enough so that
E < e−1
where the supremum is taken over all allowed input distributions. We choose R(·) to be the following distribution:
R(y) =
D W (·|x)R(·)
∞
X
=
Poisx+cE (y) · log
(112a)
(112b)
1
log E1
,
y ≥ 1.
(120)
= e−(x+cE) log
−
∞
X
(121)
1
1 − (1 + c)E
Poisx+cE (y)
y=1
(112c)
1
1
− log log − log(1 + c) + 2 + log 13
E
E
1
1
log
− (1 + c)E
+
E
1 − (1 + c)E
1
1
c2
+ E log log + (1 + c) log
. (113)
1
2
E
1−
log E1
CPE (E, c) ≤ log
Proof of (20) using Proposition 3: The last three
summands on the right-hand side of (113) are all of the form
o(1). Setting c = 0 in (113) we have
(114)
(115)
Proof of (21) using Proposition 3: From (113) we have
1
1
lim CPE (E, c) − log + log log
E↓0
E
E
≤ − log(1 + c) + 2 + log 13.
(116)
Hence
1
1
+ log log
E
E
lim lim
c→∞ E↓0
log c
− log(1 + c) + 2 + log 13
≤ lim
c→∞
log c
= −1.
1
log E1
y−1
1
− H(Y |X = x)
R(y)
y=0
then
1
1
CPE (E, 0) ≤ log − log log + 2 + log 13 + o(1)
E
E
1
1
= log − log log + O(1).
E
E

(1 + c)E 1 −
y=0
For any x ≥ 0 we have
1
−
1−2 e−1
1
e
E log <
E
1+c
4
1
144
E log
<
,
E
(1 + c)2


c)E,
1 − (1 + CPE (E, c) − log
(
· log (1 + c)E 1 −
1
log E1
1
log E1
y−1 )
− H(Y |X = x)
= e−(x+cE) log
(122)
∞
X
1
1
Poisx+cE (y)y log log
+
1 − (1 + c)E y=1
E
|
{z
}
=E[Y |X=x]=x+cE
+
∞
X
!
Poisx+cE (y)
y=1
|
{z
=1−e−(x+cE)

}


1
1

1 + log

1
(1 + c)E log E
1−
1
log E
− H(Y |X = x)
1
1
= e−(x+cE) log
+ (x + cE) log log
1 − (1 + c)E
E
+ (1 − e−(x+cE))



·
log

1
1

1 + log
1 
(1 + c)E log E
1−
log E1
− H(Y |X = x).

·
log
(123)
(124)
We can lower-bound H(Y |X = x) as
(117)
(118)
Proof of Proposition 3: Like in [2], we use the duality
bound [13] which states that, for any distribution R(·) on the
output, the channel capacity satisfies
C ≤ sup E D W (·|X)kR(·) ,
(119)
H(Y |X = x) ≥ Hb (e−(x+cE) )
= e−(x+cE)(x + cE)
+ (1 − e−(x+cE)) log
(125)
1
. (126)
1 − e−(x+cE)
Using this and (124), we upper-bound D W (·|x)R(·) as
9
≤ E + log
1
− (1 + c)E
1 − (1 + c)E
1
+ (x + cE) log log + (x + cE) log
E
D W (·|x)R(·)
1
1
+ (x + cE) log log
1 − (1 + c)E
E
+ (1 − e−(x+cE) )


≤ e−(x+cE) log
−e

·
log
−(x+cE)

1
1

1 + log

1
(1 + c)E log E
1−
1
log E
(x + cE) − (1 − e
−(x+cE)
) log
1
1 − e−(x+cE)
(127)
1
1
+ (x + cE) log log
= e−(x+cE) log
1 − (1 + c)E
E
1
− e−(x+cE) (x + cE)
+ (1 − e−(x+cE) ) log
1
| {z }
|
{z
}
1−
≤x+cE
≥cE
log E1
{z
}
|
≥0
1 − e−(x+cE)
+ (1 − e−(x+cE) ) log
(1 + c)E log E1
1
−
cE
log
≤ |e−(x+cE)
{z }
1 − (1 + c)E
≤1
|
{z
}
(128)
≥(1+c)E−cE≥0
+ (x + cE) log log
1
+ (x + cE) log
E
(1 − e−(x+cE) )
{z
}
|
+
log
=(1−e−cE )+(e−cE −e−(x+cE) )
+ (e
−cE
−e
−(x+cE)
1−
1
log E1
1 − e−(x+cE)
(1 + c)E log E1
1
− cE
1 − (1 + c)E
1
+ (x + cE) log log + (x + cE) log
E
≤ log
1
(129)
1
1−
1
log E1
1 − e−(x+cE)
(1 + c)E log E1
c
+ (1 − e−cE ) log
| {z }
(1 + c) log E1
2
|
{z
}
≥cE− c E 2
(130)
1
1−
1
log E1
+ e−cE (1 − e−x ) log
2
≤− log log
+ (1 − e−cE ) log
| {z } |
≤cE
1
E
1 − e−(x+cE)
{zcE
}
≤log
x+cE
cE
x
≤ cE
1−
1
log E1
1 − e−(x+cE)
+ e−cE (1 − e−x ) log
(1 + c)E log E1
c2
1
− cE − E 2 log log + x.
2
E
(132)
Using (119) and (132) together with constraint (3) we have
C(E, c)
≤ sup E D W (·|X)kR(·)
1
≤ E + log
− (1 + c)E
1 − (1 + c)E
1
+ (1 + c)E log log + (1 + c)E log
E
(133)
1
1−
1
log E1
1
c2 2
− cE − E log log + E
2
E
1 − e−(X+cE)
−cE
−X
+e
sup E (1 − e ) log
(134)
(1 + c)E log E1
1
= E log log + 2E
E
1
− (1 + c)E
+ log
1 − (1 + c)E
1
1
c2
+ E 2 log log + (1 + c)E log
1
2
E
1−
log E1
1 − e−(X+cE)
−cE
−X
+e
sup E (1 − e ) log
. (135)
(1 + c)E log E1
Hence the maximum achievable photon efficiency satisfies
1 − e−(x+cE)
) log
(1 + c)E log E1
1 − e−(x+cE)
+ (1 − e−cE ) log
(1 + c)E log E1
1
= E + log
− (1 + c)E
1 − (1 + c)E
1
+ (x + cE) log log + (x + cE) log
E
1
(131)
CPE (E, c)
C(E, c)
(136)
=
E
1
1
1
≤ log log + 2 +
log
− (1 + c)E
E
E
1 − (1 + c)E
2
1
1
c
+ E log log + (1 + c) log
1
2
E
1−
log E1
1 − e−(X+cE)
e−cE
−X
sup E (1 − e ) log
. (137)
+
E
(1 + c)E log E1
We next upper-bound the expectation on the right-hand side
of (137) as follows
1 − e−(X+cE)
E (1 − e−X ) log
(1 + c)E log E1
1 − e−X
1 − e−(X+cE)
=E X·
log
(138)
X
(1 + c)E log E1
X + cE
1 − e−X
log
≤E X·
(139)
X
(1 + c)E log E1
≤ E · sup φ(x)
(140)
10
where
Plugging (149), (153) and (154) into (143) we have
φ(x) ,
1−e
x
−x
log
x + cE
,
(1 + c)E log E1
x ≥ 0.
(141)
We claim that, if E is small enough to satisfy all of the
conditions in (112), then
12
sup φ(x) = φ(x ) for some x ≤
.
log E1
∗
∗
1
x + cE
1
dφ(x)
≤ − log
(155)
1 + x
dx
6
(1 + c)E log E
1
1
1
1
≤−
+ log(1 + c) − log
log x
| {z }
6
6
6
E
≥log 12−log log
(142)
+
To show this, we compute the derivative of φ(x) (with
respect to x) as
−x
−1
dφ(x)
< 0,
dx
12
x>
.
log E1
(145)
(146)
Note that this case need not be considered if E ≥ e−12 .
In this case the logarithmic term in the derivative is positive:
log
x
x + cE
1 ≥ log
(1 + c)E log E
(1 + c)E log E1
12
> log
2
(1 + c)E log E1
> 0,
(149)
where the last step follows because (112a) and (112c) together
imply
2
1
1
12 e− 2
E log
<
.
(150)
E
1+c
Next, using Taylor’s theorem with Cauchy remainder, we have
′
1 1
(3 − x′ ) e−x 2
(1 + x) e−x − 1
=
−
+
x
−
x
x2
2 3
24{z
|
}
(151)
≥0
1 1
x
≤ − + |{z}
2 3
(152)
≤1
1
≤− ,
(153)
6
where x′ ∈ [0, x]. The underbrace in (151) follows because
x′ ≤ x < 1. We also have
x
1
1 − e−x
≤
≤ .
x(x + cE)
x(x + cE)
x
(154)
(157)
(158)
(159)
where the last step follows from (112c). We have now shown
dφ(x)
< 0,
dx
12
< x < 1.
log E1
(160)
Case 2. Consider
x ≥ 1.
(161)
In this case the logarithmic term in the derivative is still
positive:
1
x + cE
> 0,
1 ≥ log
(1 + c)E log E
(1 + c)E log E1
(162)
where the last step follows from (112b). We bound the other
terms as
(1 + x) e−x − 1
1 − 2 e−1
≤−
(163)
2
x
x2
which is negative, and
(147)
(148)
1
E
< 0,
log
12
< x < 1.
log E1
(156)
1
1+c 1
1
1
1
log
+ log log −
log
6
12
3
E
12
E
4 !
1
1
(1 + c)2
=
log
E log
12
144
E
We do this separately for two cases.
Case 1. Consider
1
x
|{z}
≤
log
dφ(x)
= 0.
(144)
dx
To prove (142), it thus suffices to show that, under the
assumptions (112),
1
1
log log +
6
E
1
≤ 12
log
−x
1−e
x + cE
+
x(x + cE)
(1 + c)E log E1
(143)
which exists and is continuous for all x > 0. Further note
that it is negative for large enough x, and is positive for small
enough x. Hence x∗ must be achieved at a point where
(1 + x) e
dφ(x)
=
dx
x2
1
E
1 − e−x
1
≤ 2.
x(x + cE)
x
(164)
Using (162), (163) and (164) together with (143) we have


dφ(x)
1
1 − 2 e−1 

≤−
log(x + cE) + log
1 
{z
}
|
dx
x2
(1 + c)E log E
≥0
1
+ 2
x
1 − 2 e−1
1
≤−
log
x2
(1 + c)E log E1
1 − 2 e−1
1
+
·
x2
1 − 2 e−1
=−
< 0,
(165)
(166)
− 1−21e−1
e
1 − 2 e−1
log
2
x
(1 + c)E log E1
(167)
(168)
where the last step follows from (112b). Hence we have shown
dφ(x)
< 0, x ≥ 1.
(169)
dx
Combining (160) and (169) proves (142) for all E satisfying (112).
11
We now proceed to upper-bound sup φ(x):
sup φ(x) = φ(x∗ )
∗
1 − e−x
x∗ + cE
=
log
∗
(1 + c)E log E1
| x{z }
(140), and is given by
(170)
(171)
≤1
x∗ + cE
(1 + c)E log E1


12


+
cE




log E1
≤ max 0, log
.

(1 + c)E log E1 




≤ max 0, log
(172)
1
.
log E1
where φ(·) is given in (141) and sup φ(x) is computed
numerically. The on-off-keying lower bound is obtained by
computing the mutual information of the channel with “on”
signal equaling η, and with the receiver ignoring multiple
detected photons. This is given by
(174)
CPE-OOK (E, c)
E
1
E
Hb (q) − · Hb (e−η−cE ) − 1 −
Hb (e−cE ) ,
≥
E
η
η
(181)
Here (176) follows because conditions (112a) and (112c)
imply (150).
Combining (137), (140) and (177), and noting e−cE ≤ 1
(together with the fact that the right-hand side of (177) is
positive), we obtain
CPE (E, c)
1
1
1
log
− (1 + c)E
≤ log log + 2 +
E
E
1 − (1 + c)E
c2
1
1
+ E log log + (1 + c) log
1
2
E
1−
log E1
1
1
+ log − 2 log log − log(1 + c) + log 13 (178)
E
E
1
1
= log − log log − log(1 + c) + 2 + log 13
E
E
1
1
log
− (1 + c)E
+
E
1 − (1 + c)E
c2
1
1
+ E log log + (1 + c) log
.
(179)
1
2
E
1−
log E1
AND
(180)
(173)
We can now continue the chain of inequalities (173) as
)
(
13
(175)
sup φ(x) ≤ max 0, log
2
(1 + c)E log E1
13
= log
(176)
2
(1 + c)E log E1
1
1
= log − 2 log log − log(1 + c) + log 13. (177)
E
E
VII. N UMERICAL C OMPARISON
R EMARKS
1
1
1
log
− (1 + c)E
≤ log log + 2 +
E
E
1 − (1 + c)E
1
1
c2
+ E log log + (1 + c) log
1
2
E
1−
log E1
+ e−cE sup φ(x),
Now note that, due to (112b), we have
cE ≤
CPE (E, c)
C ONCLUDING
We numerically compare the approximation (23) with
nonasymptotic upper and lower bounds on photon efficiency.
Specifically, the plotted upper bound is based on (137) and
where Hb (·) denotes the binary entropy function
Hb (a) = a log
1
1
+ (1 − a) log
,
a
1−a
a ∈ [0, 1],
(182)
and where
E
E −η−cE
e−cE .
+ 1−
q, e
η
η
(183)
The simple-PPM lower bound is computed using (51) and
(57):
1
bp0
CPE-PPM (E, c) ≥
p0 log
η
p0 + (b − 1)p1
bp1
. (184)
+ (b − 1)p1 log
p0 + (b − 1)p1
The soft-decision-PPM lower bound is computed using (96)
and is given by
CPE-PPM(SD) (E, c)
1
bp0
≥
p0 log
η
p0 + (b − 1)p1
bp1
p0 + (b − 1)p1
bp2
+ (b − 1)p2 log
2p2 + (b − 2)p3
bp3
(b − 1)(b − 2)
. (185)
p3 log
+
2
2p2 + (b − 2)p3
+ (b − 1)p1 log
In all these lower bounds, the parameters b and η are chosen as
in (26) and (28), and p0 , p1 , p2 , p3 , and p4 are given in (41),
(47), (84), (87), and (89), respectively. We plot these bounds
for c = 0.1, c = 1, and c = 10 in Figure 1.
The figure shows the following.
• The approximation (23) is reasonably accurate for small
enough E.
• The on-off-keying and the soft-decision-PPM bounds are
consistently close to the approximation (23).
12
As c increases, the simple-PPM bound deviates significantly from the other lower bounds, and hence also from
the actual value of CPE (E, c).
The capacity bounds as well as the asymptotic results in
this paper provide insights to how technical restrictions and
device or channel imperfections influence the communication
rates in optical channels in the wideband regime. Our results
and techniques can also be applied to secret communication
and key distribution over optical channels [14]. It would be
interesting to see if these are further extendable, e.g., to
multiple-mode optical channels and optical networks.
As we have demonstrated, PPM is nearly optimal on the
Poisson channel in the high-photon-efficiency regime. When
c is small, the simple-PPM super channel has high erasure
probability but low “error” (by which we mean the receiver
detects a single pulse at a position that is different from the
transmitted signal) probability. In this case, Reed-Solomon
codes can perform rather close to the theoretical limit. However, when c > 1, Reed-Solomon codes can no longer achieve
any positive rates on this channel. Nevertheless, we believe
that, for c > 1, PPM still has its advantages over on-off
keying in terms of code design. This is because the optimal
input distribution for (both simple and soft-decision) PPM
is uniform, whereas the optimal input distribution for on-off
keying for this channel is highly skewed. The uniformity of
PPM inputs allows the usage of more structured codes, in
particular linear codes. One possible direction, for instance,
is to employ the idea of multilevel codes [15], [16] on this
channel.
•
20
Photon Efficiency (nats/photon)
Upper bound
18
Approximation (23)
On-off keying
16
Soft-decision PPM
14
Simple PPM
12
10
8
6
4
2
0
−8
10
−7
10
−6
10
−5
E
−4
10
10
−3
10
(a) Case c = 0.1. Note that the lowest three curves almost
overlap.
Photon Efficiency (nats/photon)
20
Upper bound
18
Approximation (23)
On-off keying
16
Soft-decision PPM
14
Simple PPM
12
10
8
6
4
2
0
−8
10
−7
10
−6
10
−5
E
−4
10
10
−3
10
(b) Case c = 1. The on-off keying and soft-decision PPM
curves overlap.
A PPENDIX
In this appendix we provide a proof for the converse
part of (8) that is much simpler than the one given in [2].
Using the same arguments as in [2], we know that C(E, c) is
monotonically increasing in c. Hence it suffices to show
C(E, 0)
≤ 1.
(186)
lim
E↓0 E log 1
E
This can be verified as follows:
C(E, 0) = max I(X; Y )
E[X]≤E
Photon Efficiency (nats/photon)
20
Upper bound
18
On-off keying
16
(188)
= max H(Y )
(189)
= (1 + E) log(1 + E) − E log E,
(190)
E[Y ]≤E
Soft-decision PPM
14
≤ max H(Y )
E[X]≤E
Approximation (23)
Simple PPM
(187)
12
which yields the desired asymptotic bound. Here, (190) follows from the well-known fact that, for a nonnegative discrete
random variable under a first-moment constraint, maximum
entropy is achieved by the geometric distribution, and is given
by the right-hand side of (190). Note that the right-hand
side of (190) is exactly the capacity of the pure-loss bosonic
channel [7].
10
8
6
4
2
0
−8
10
−7
10
−6
10
−5
E
10
−4
10
−3
10
(c) Case c = 10.
Fig. 1. Comparing the approximation (23) to nonasymptotic upper bound,
and to lower bounds for simple PPM, soft-decision PPM, and on-off keying
for three cases: c = 0.1, c = 1, and c = 10.
ACKNOWLEDGMENTS
The authors thank Hye Won Chung, Yuval Kochman, Amos
Lapidoth, Arya Mazumdar, and Lizhong Zheng for helpful
discussions, and the anonymous reviewers for their careful
reading of the manuscript.
13
R EFERENCES
[1] J. H. Shapiro, “The quantum theory of optical communications,” IEEE
J. Select. Top. Quantum Electron., vol. 15, no. 6, pp. 1547–1569, 2009.
[2] A. Lapidoth, J. H. Shapiro, V. Venkatesan, and L. Wang, “The discretetime Poisson channel at low input powers,” IEEE Trans. Inform. Theory,
vol. 57, no. 6, pp. 3260–3272, June 2011.
[3] J. Pierce, “Optical channels: practical limits with photon counting,” IEEE
Trans. Commun., vol. 26, no. 12, pp. 1819–1821, Dec. 1978.
[4] J. L. Massey, “Capacity, cutoff rate, and coding for direct-detection
optical channel,” IEEE Trans. Commun., vol. 29, no. 11, pp. 1615–1621,
Nov. 1981.
[5] S. Shamai (Shitz), “Capacity of a pulse amplitude modulated direct detection photon channel,” in Proc. IEE, vol. 137, pt. I (Communications,
Speech and Vision), no. 6, Dec. 1990, pp. 424–430.
[6] A. Lapidoth and S. M. Moser, “On the capacity of the discrete-time
Poisson channel,” IEEE Trans. Inform. Theory, vol. 55, no. 1, pp. 303–
322, Jan. 2009.
[7] V. Giovannetti, S. Guha, S. Lloyd, L. Maccone, J. H. Shapiro, and
H. P. Yuen, “Classical capacity of the lossy bosonic channel: the exact
solution,” Phys. Rev. Lett., vol. 92, no. 2, p. 027902(4), 2004.
[8] R. G. Gallager, “Energy limited channels: Coding, multiaccess, and
spread spectrum,” Laboratory for Information and Decision Systems,
MIT, Tech. Rep. 1714, Nov. 1987.
[9] S. Verdú, “On channel capacity per unit cost,” IEEE Trans. Inform.
Theory, vol. 36, pp. 1019–1030, Sept. 1990.
[10] H. W. Chung, S. Guha, and L. Zheng, “On capacity of optical channels
with coherent detection,” in Proc. IEEE Int. Symp. Inform. Theory, Saint
Petersburg, Russia, July 31–August 5 2011.
[11] Y. Kochman and G. W. Wornell, “On high-efficiency optical communication and key distribution,” in Proc. Inf. Theory and Appl. Workshop,
San Diego, CA, USA, Feb. 5–10 2012.
[12] S. Dolinar, K. M. Birnbaum, B. M. Erkmen, and B. Moision, “On approaching the ultimate limit of photon-efficient and bandwidth efficient
optical communication,” in Int. Conf. Space Optical Sys.and App., Santa
Monica, CA, USA, May 11—13 2011.
[13] A. Lapidoth and S. M. Moser, “Capacity bounds via duality with
applications to multiple-antenna systems on flat fading channels,” IEEE
Trans. Inform. Theory, vol. 49, no. 10, pp. 2426–2467, Oct. 2003.
[14] Y. Kochman, L. Wang, and G. W. Wornell, “Toward photon-efficient key
distribution over optical channels,” 2013, subm. to IEEE Trans. Inform.
Theory.
[15] H. Imai and S. Hirakawa, “A new multilevel coding method using error
correcting codes,” IEEE Trans. Inform. Theory, vol. 23, pp. 371–377,
1977.
[16] U. Wachsmann, R. F. H. Fischer, and J. B. Huber, “Multilevel codes:
Theoretical concepts and practical design rules,” IEEE Trans. Inform.
Theory, vol. 45, no. 5, pp. 1361–1391, July 1999.